Self-Distilling Future Video Representations for World Action Models